znshje / 3DTeethSeg22_IGIP

3DTeethSeg22_IGIP

Getting started

Training

First, install the python requirements by:

pip install -r requirements.txt

Next, some operators should be compiled in code/pointnet2. Run the following command to compile:

cd code/pointnet2/src
pip install -e .

You should train all the models in code/models. The training and testing config is written in code/config/config.yml. After adjusting the configuration files, just run the python scripts in code/models directly to start training.

# Training steps
# Change any args you want
export CUDA_VISIBLE_DEVICES=0 && python code/models/teeth_gingival_seg.py
export CUDA_VISIBLE_DEVICES=0 && python code/models/centroids_prediction.py
export CUDA_VISIBLE_DEVICES=0 && python code/models/patch_segmentation.py
export CUDA_VISIBLE_DEVICES=0 && python code/models/teeth_classification.py

You should better keep the default learning rate (lr) and epoch nums (n-epochs), they work good for us.

Testing

There's no testing code in this repository. BUT, you can do inferences by code/process.py, this script is based on the challenge's process.py.

GPU and memory consumption

Here we show the GPU memory and host memory consumption when setting the specific batch size for reference. The data is obtained on a machine with one NVIDIA RTX 3090. The batch size can be adjusted according to your device.

Stage	Batch size	GPU memory	Host memory
teeth_gingival_separation	16	~20GB	~4GB
centroids_prediction	16	~16GB	~4GB
patch_segmentaiton	16	~21GB	~5GB
teeth_classification	64	~16GB	~5GB

Suggested environment

The following environment is tested ok.

• Linux arch 5.15.74-1-lts
• GCC 9.3.0
• Python 3.8
• PyTorch 1.10.0
• CUDA 11.7

Support

I do some code clean up based on my original codes, and I didn't re-train all the models. They should work fine, but there's no guarantee. If you have any problems, please e-mail me at [e-mail].